mixed/README.md

metadata

license: other
license_name: nvidia-open-model-license
license_link: https://huggingface.co/nvidia/personaplex-7b-v1/resolve/main/LICENSE
language:
  - en
library_name: onnxruntime
tags:
  - personaplex
  - speech-to-speech
  - full-duplex
  - voice-agent
  - onnx
  - moshi-architecture
base_model: nvidia/personaplex-7b-v1

PersonaPlex 7B ONNX -- mixed variant

Pareto winner on quality + VRAM. INT8-dynamic temporal + FP16 depformer + FP32 mimi.

This is one of four ONNX-quantized bundles of NVIDIA's PersonaPlex 7B -- a full-duplex speech-to-speech model on Kyutai's Moshi architecture. The full collection is at soniqo/PersonaPlex-7B-ONNX.

Quick reference (measured on RTX 5090, VARF2 voice, "helpful" prompt, 50 frames)

	This bundle
Disk	~11 GB
Host RAM	7.9 GB
VRAM	6.6 GB
RTF (12.5 Hz)	3.5x (1.0 = realtime)
Quality	cos 0.990448 vs FP32 reference

Best quality at low VRAM. RTF is the cost of INT8 dynamic Q/DQ.

All four variants

Variant	Disk	Host RAM	VRAM	RTF	hidden cos	Best for
fp16	~17 GB	1.5 GB	18.3 GB	5.3x	1.0000	Near-perfect quality
mixed	~11 GB	7.9 GB	6.6 GB	3.5x	0.9904	Best quality at low VRAM ← (this)
int8-nb-dep_gint8 ⭐	~9.4 GB	1.4 GB	12.1 GB	1.12x	0.9977	Best balance: near realtime RTF, low host RAM, excellent quality
int4-nb-dep_gint8	~7.6 GB	1.4 GB	9.6 GB	1.12x	0.8774	Smallest viable bundle

The int8-nb-dep_gint8 variant is the recommended ship default: best RTF (1.12x, near realtime), low host RAM (1.4 GB), and excellent quality (cos 0.998).

The mixed variant wins on quality + VRAM combined but its RTF is 3.5x: the INT8 dynamic-quantize pattern adds 384 CPU↔GPU Memcpy bridges which block CUDA Graph capture. Pick this when VRAM is the binding constraint.

How to use

This bundle is consumed by the speech-core C++ inference runtime:

# Download (from speech-core repo root)
PERSONAPLEX_VARIANT=mixed scripts/download_personaplex_onnx.sh

# Run
build/Release/run_personaplex scripts/personaplex-mixed 50 audio.wav VARF2

Or load directly with ONNX Runtime in any language:

import onnxruntime as ort
sess = ort.InferenceSession("temporal_step.onnx", providers=["CUDAExecutionProvider"])
# inputs: text_token [1,1] int64, audio_tokens [1,16] int64,
#         past_k_all [32,1,32,T,128] float, past_v_all [32,1,32,T,128] float
# outputs: hidden [1,1,4096] float, new_k_all, new_v_all

Files in this variant

File	Purpose
mimi_encoder.onnx(+.data)	24 kHz PCM -> 16 audio codebooks @ 12.5 Hz
mimi_decoder.onnx(+.data)	16 audio codebooks @ 12.5 Hz -> 24 kHz PCM
temporal_step.onnx(+.data)	One frame of 32-layer 7B temporal transformer, explicit KV-cache I/O
depformer_step.onnx(+.data)	One inner step of 6-layer depformer, 16 codebook steps per frame
tokenizer_spm_32k_3.model	SentencePiece text tokenizer
voices/<name>.bin	18 voice prompts (NATF*, NATM*, VARF*, VARM*)
system_prompts.bin	Pre-tokenized "helpful" / "expert" / "warm" / "direct" prompts
config.json	Architecture + precision + measured-metrics metadata

Related

• soniqo/speech-core -- C++ runtime that consumes this bundle, with the OnnxPersonaPlex wrapper and CUDA EP routing
• soniqo.audio -- the project site

License

Same as the upstream NVIDIA PersonaPlex 7B (NVIDIA Open Model License).